Point cloud target detection method and device based on severe weather rendering and upsampling

Abstract

The invention relates to a point cloud target detection method and device based on severe weather rendering and up-sampling. The method comprises the following steps: S1, obtaining point cloud data collected by a laser radar; s2, performing severe weather rendering on the point cloud data to generate simulated severe weather point cloud; s3, carrying out severe weather noisy point removal and optimization on the rendered point cloud data through a deep learning method; s4, inputting the denoised point cloud data into an up-sampling network, and complementing effective point loss caused by severe weather rendering; and S5, inputting the point cloud data processed by the up-sampling network into the three-dimensional target detection model, carrying out point cloud target detection, generating a target detection prediction result, and completing final identification and positioning. According to the method, the three-dimensional position and category of the target in severe weather can be accurately detected, and the robustness and real-time performance of the detection model are effectively improved. The adaptive capacity of the method to road scenes is higher, and meanwhile, the computing resource consumption of the algorithm is remarkably reduced.

Description

Technical Field

[0001] The present disclosure relates to the technical field of target detection, and in particular, to a point cloud target detection method and device based on severe weather rendering and upsampling. Background Art

[0002] With the maturity of autonomous driving technology, smart cars equipped with different levels of autonomous driving functions have gradually entered the public's field of vision. Three-dimensional object detection is one of the important tasks of autonomous driving technology. At present, three-dimensional object detection methods are mainly used to detect road scenes in clear weather. However, in severe weather conditions, factors such as moving particles in the air (such as raindrops, snowflakes) or suspended particles (such as fog droplets, dust haze) and wet ground will have a significant impact on the lidar point cloud, resulting in serious performance loss of existing three-dimensional detection algorithms. Researchers regard rain and snow as noise in the algorithm and conduct research on rain and snow removal. Traditional denoising methods include statistical outlier removal (SOR), dynamic statistical outlier removal (DSOR) and low-intensity outlier removal (LIOR). These methods construct filtering models by analyzing the intensity and distance information of point clouds. In addition to denoising algorithms, some studies also deal with sparse point problems by improving the quality of point clouds. For example, SPG generates semantic points to restore the shape of three-dimensional objects. Considering the problem of incomplete point clouds caused by signal loss, pseudo point clouds generated by images are used to supplement point clouds. However, traditional methods lack adaptability to road scenes, which limits the real-time performance of autonomous driving. Summary of the invention

[0003] In order to overcome the problem that traditional three-dimensional target detection methods lack adaptability to road scenes and limit the real-time performance of autonomous driving, the present invention provides a point cloud target detection method and device based on severe weather rendering and upsampling.

[0004] In a first aspect, the present disclosure provides a point cloud target detection method based on severe weather rendering and upsampling, comprising:

[0005] S1. Obtain point cloud data collected by the laser radar;

[0006] S2. Perform bad weather rendering on the point cloud data to generate a simulated bad weather point cloud;

[0007] S3. Remove and optimize the bad weather noise from the rendered point cloud data using deep learning methods;

[0008] S4. Input the denoised point cloud data into the upsampling network to complete the loss of valid points caused by bad weather rendering;

[0009] S5. Input the point cloud data processed by the upsampling network into the 3D target detection model, perform point cloud target detection, generate target detection prediction results, and complete the final recognition and positioning.

[0010] Optionally, in step S1, 60% of the constructed snow and fog severe weather data set is used as a training set and 40% is used as a test set.

[0011] Optionally, when rendering foggy weather based on point cloud, step S2 includes:

[0012] S2-1. Input the point cloud data collected by the LiDAR under a clear sky, where each point contains the spatial position (x, y, z);

[0013] S2-2. Calculate the original position R0 of each point:

[0014]

[0015] S2-3. Calculate the hard target item of the point using R0 and the attenuation coefficient of the fog;

[0016] S2-4. Calculate the soft target term using R0, the initial scattering power of the fog, the backscattering system of the fog and the current signal propagation distance,

[0017] S2-5. Compare and if the soft target item of the point is greater than the hard target item, update the spatial position of the point and then output it; otherwise, directly output the position information of the point;

[0018] S2-6. Use the output spatial position information as simulation data under foggy conditions for training and testing of three-dimensional target detection.

[0019] Optionally, in step S2-3, the hard target item The formula is:

[0020] Where: α represents the attenuation coefficient of fog, P R，clear (R) is the received signal power in clear weather;

[0021] In step S2-4, the soft target item The formula is:

[0022] Where: CAP0 represents the initial scattering power of the fog, β represents the backscattering system of the fog, and R represents the distance of current signal propagation.

[0023] Optionally, in step S2-5, the formula for updating the spatial position of the point is: x′=s·n·x, y′=s·n·y, z′=s·n·z, otherwise the point is directly output, where: s is the scaling factor, and n is the noise factor.

[0024] Optionally, when rendering snow weather based on point cloud, step S2 includes:

[0025] S2-7. Input the point cloud data collected by the radar under a clear sky, where each point contains the spatial position (x, y, z);

[0026] S2-8. Generate snow particles based on the snowfall rate and the parameters of the radar sensor, and the snow particles are distributed around the radar sensor and conform to a random distribution;

[0027] S2-9. Calculate the original distance R0 from the point in the scene scanned by the laser radar to the laser radar sensor, determine the number of snow particles intersecting the light beam on the scanning path, and gradually calculate the snow particle occlusion and scattering effects;

[0028] S2-10. Calculate the hard target at each point, calculate the sum of the received power of the hard target at that point and all the snow particles in the current beam path at that point, get the total received power, take the maximum value of the received power at that point, determine the reflected power and the corresponding position according to the maximum value of the received power, and calculate the received power of the hard target

[0029] Where: CAP0 represents the initial scattering power of fog, α is the incident angle, and ·ρ0 represents the reflectivity of the target object;

[0030] For receiving power: Where: δ represents the overlap between the receiver and the transmitter. ρ s represents the reflectivity of snow particles, c represents the light beam, τ H represents the half-power pulse width, θ j It indicates the part of the laser beam opening angle reflected by snow particles;

[0031] Calculate the sum of the received power of the hard target and all snow particles to get the total received power P R，snow :

[0032] S2-11. Compare the peak value of the reflected power at the point and the size of the hard target at the point. If the peak value of the reflected power is greater than the hard target, the position of the point is updated and then output; otherwise, the position information of the point is directly output;

[0033] S2-12. Use the output spatial position information as simulation data under snowy conditions for training and testing of three-dimensional target detection.

[0034] Optionally, step S3 includes:

[0035] S3-1. Input the rendered point cloud data;

[0036] S3-2. Convert the point cloud data into a spherical projection image to enhance the spatial feature expression of the point cloud;

[0037] S3-3. Use the neighborhood feature construction method based on KNN convolution to extract the local geometric features of the spherical projection image;

[0038] S3-4. Capture the temporal information of the point cloud from the time dimension, combine the spatial features of adjacent points, and generate local high-dimensional features;

[0039] S3-5. Through multi-resolution convolution, feature enhancement and noise reduction strategies, severe weather noise is removed with high precision;

[0040] S3-6. Extract valid points and reconstruct 3D point cloud data, significantly improving the integrity of the point cloud and the accuracy and reliability of target detection.

[0041] S3-7. Output the denoised point cloud data.

[0042] Optionally, step S4 includes:

[0043] S4-1. Perform multi-scale local feature extraction and modeling on the denoised point cloud data;

[0044] S4-2. Perform generator-discriminator adversarial training on the modeled point cloud data;

[0045] S4-3. Perform high-quality processing on the trained point cloud data by designing multiple loss functions.

[0046] Optionally, step S5 includes:

[0047] S5-1. voxelize the point cloud data processed by the multiple loss functions to divide it into multiple small voxels;

[0048] S5-2. Feature encoding of points in each voxel is performed by means of a multi-layer perceptron to obtain voxel features;

[0049] S5-3. Perform convolution operation on voxel features through 3D convolutional neural network to extract high-level semantic features;

[0050] S5-4. Generate 3D candidate boxes through region proposal network using 3D feature map;

[0051] S5-5. Select some key points from the point cloud data processed by the multiple loss functions, pool relevant features from the voxel features around the key points, aggregate the voxel features to the key points, and obtain key point level features;

[0052] S5-6. Fuse high-level semantic features with key point-level features, and use the fused features to refine and classify the 3D candidate boxes, determine the types of target objects contained therein, and finally obtain an accurate 3D target detection box.

[0053] According to a second aspect of an embodiment of the present disclosure, a point cloud target detection device based on snow and fog severe weather is provided, comprising:

[0054] An acquisition module, configured to acquire point cloud data collected by a laser radar;

[0055] a rendering module configured to perform severe weather rendering on the point cloud data to generate a simulated severe weather point cloud;

[0056] A denoising module configured to remove and optimize severe weather noise from rendered point cloud data using a deep learning approach;

[0057] The data completion module is configured to input the denoised point cloud data into the upsampling network to complete the loss of valid points caused by bad weather rendering;

[0058] The target detection module is configured to input the point cloud data processed by the upsampling network into the three-dimensional target detection model, perform point cloud target detection, generate target detection prediction results, and complete the final recognition and positioning.

[0059] The technical solution provided by the embodiments of the present disclosure may have the following beneficial effects:

[0060] Based on the existing point cloud detection model, the point cloud rendering snow and fog weather algorithm is combined to render the bad weather effect on the existing highly compatible and standardized clear weather data set. The snow and fog noise points of the rendered data set are removed, and the results are introduced into the point cloud upsampling network to complete the valid points of the objects lost due to bad weather. The three-dimensional target detection in bad weather is recognized and analyzed by the neural network, which realizes the accurate detection of the target position and category, and effectively improves the robustness and real-time performance of the detection model, which ensures the strong adaptability of the recognition of road scenes and reduces the consumption of a large amount of computing resources by the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a first flow chart of a point cloud target detection method based on snow and foggy bad weather according to an exemplary embodiment of the present disclosure.

[0062] Figure 2 It is a second flow chart of a point cloud target detection method based on snow and foggy bad weather according to an exemplary embodiment of the present disclosure.

[0063] Figure 3It is a third flow chart of a point cloud target detection method based on snow and foggy bad weather according to an exemplary embodiment of the present disclosure. DETAILED DESCRIPTION

[0064] The specific implementation of the present disclosure is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described herein is only used to illustrate and explain the present disclosure, and is not used to limit the present disclosure.

[0065] In this disclosure, unless otherwise stated, directional words such as "upper, lower, left, right" are used for the convenience of description and are defined according to the drawing direction of the corresponding drawings, and "inside, outside" are defined according to the contours of the corresponding parts. Terms such as "first, second" and the like used in this disclosure are used to distinguish one element from another and do not have order or importance. In addition, when the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements.

[0066] Before further describing the embodiments of the present application in detail, the nouns and terms involved in the embodiments of the present application are explained. The nouns and terms involved in the embodiments of the present application are subject to the following interpretations.

[0067] 1) Deep learning is a machine learning technology based on artificial neural networks. It automatically extracts features through a multi-layer network structure to complete complex tasks such as object detection, speech recognition, and language translation. The core of deep learning is its ability to learn and extract useful features from large amounts of data without manually writing rules or knowledge.

[0068] 2) The upsampling network is a point cloud upsampling method based on the generative adversarial network (GAN). The core idea is to learn the distribution characteristics of the local geometric structure of the point cloud through adversarial training, so as to interpolate and generate more realistic completion points on the sparse point cloud.

[0069] The embodiments of the present application provide a point cloud target detection method, device, electronic device, computer-readable storage medium and computer program product based on severe weather rendering and upsampling, which can improve the robustness and real-time performance of point cloud target detection.

[0070] The following describes an exemplary application of an electronic device provided by an embodiment of the present application. The electronic device provided by an embodiment of the present application can implement a terminal device, such as a laptop, a tablet computer, a desktop computer, a set-top box, a smart TV, a mobile device (e.g., a mobile phone, a portable music player, a personal digital assistant, a dedicated messaging device, a portable gaming device), a vehicle terminal, a virtual reality (VR) device, an augmented reality (AR) device, and other types of user terminals, and can also be implemented as a server. Below, an exemplary application when the electronic device is implemented as a terminal device or a server will be described.

[0071] See also Figure 1 and Figure 2 In a first aspect, the present disclosure provides a point cloud target detection method, comprising:

[0072] S1. Obtain point cloud data collected by LiDAR to provide a basic data set for subsequent processing.

[0073] S2. Perform bad weather rendering on the point cloud data to generate simulated bad weather point clouds to enhance the system's ability to adapt to bad weather.

[0074] S3. Use deep learning methods to remove and optimize the noise points of severe weather on the rendered point cloud data, and remove unnecessary interference data caused by weather simulation.

[0075] S4. Input the denoised point cloud data into the upsampling network to complete the valid points lost due to bad weather rendering and restore the integrity of the target object.

[0076] S5. Input the point cloud data processed by the upsampling network into the 3D target detection model, perform point cloud target detection, generate target detection prediction results, and complete the final recognition and positioning.

[0077] It is understandable that based on the existing point cloud detection model, the existing highly compatible and standardized clear weather data set is rendered with bad weather effects in combination with the point cloud rendering snow and fog weather algorithm. The rendered data set is subjected to snow and fog noise removal, and its results are introduced into the point cloud upsampling network to complete the effective points of the objects lost due to bad weather. The three-dimensional target detection in bad weather is recognized and analyzed by the neural network, which realizes the accurate detection of the target position and category, and effectively improves the robustness and real-time performance of the detection model, that is, it ensures the strong adaptability of the recognition of road scenes, while reducing the algorithm's consumption of a large amount of computing resources.

[0078] In one embodiment, in step S1, 60% of the constructed snow fog bad weather data set is used as a training set and 40% is used as a test set, so as to ensure that the model can effectively identify targets under snow fog bad weather conditions. Of course, in other embodiments, 70% of the constructed snow fog bad weather data set is used as a training set and 30% is used as a test set.

[0079] See also Figure 2 In one embodiment, when rendering foggy weather based on point cloud, step S2 includes:

[0080] S2-1. Input the point cloud data collected by the LiDAR under a sunny sky, where each point contains a spatial position (x, y, z).

[0081] S2-2. Calculate the original position R0 of each point:

[0082] S2-3. Calculate the hard target item of the point using R0 and the attenuation coefficient of the fog.

[0083] S2-4. Calculate the soft target term using R0, the initial scattering power of the fog, the backscattering system of the fog and the distance of the current signal propagation.

[0084] S2-5. And compare, if the soft target item of the point is greater than the hard target item, then update the spatial position of the point and output it, otherwise, directly output the position information of the point.

[0085] S2-6. Use the output spatial position information as simulation data under foggy conditions for training and testing of three-dimensional target detection.

[0086] It can be understood that the point cloud target detection method uses sunny point cloud data as the basis, and calculates the "hard target item" and "soft target item" of each point according to parameters such as the fog attenuation coefficient, initial scattering power, backscattering system and signal propagation distance, simulating the strong and weak signals of the points in the foggy lighting environment. If the soft target item is greater than the hard target item, the spatial position of the point is updated to simulate the actual position of the point cloud after foggy weather, and finally outputs the simulated foggy point cloud data, which can be used for training and testing of three-dimensional target detection.

[0087] In one embodiment, in step S2-3, the hard target item The formula is:

[0088] Where: α represents the attenuation coefficient of fog, P R，clear (R) is the received signal power in clear weather;

[0089] In step S2-4, the soft target item The formula is:

[0090] Where: CAP0 represents the initial scattering power of the fog, β represents the backscattering system of the fog, and R represents the distance of current signal propagation.

[0091] In one embodiment, in step S2-5, the formula for updating the spatial position of the point is: x′=s·n·x, y′=s·n·y, x′=s·n·x otherwise the point is directly output, wherein: s is the scaling factor, and n is the noise factor.

[0092] See also Figure 3 In one embodiment, when rendering snow weather based on point cloud, step S2 includes:

[0093] S2-7. Input the point cloud data collected by the radar under a clear sky, where each point contains the spatial position (x, y, z);

[0094] S2-8. Generate snow particles based on the snowfall rate and the parameters of the radar sensor, and the snow particles are distributed around the radar sensor and conform to a random distribution;

[0095] S2-9. Calculate the original distance R0 from the point in the scene scanned by the laser radar to the laser radar sensor, determine the number of snow particles intersecting the light beam on the scanning path, and gradually calculate the snow particle occlusion and scattering effects;

[0096] S2-10. Calculate the hard target at each point, calculate the sum of the received power of the hard target at that point and all the snow particles in the current beam path at that point, get the total received power, take the maximum value of the received power at that point, determine the reflected power and the corresponding position according to the maximum value of the received power, and calculate the received power of the hard target

[0097] Where: CAP0 represents the initial scattering power of fog, α is the incident angle, and ·ρ0 represents the reflectivity of the target object;

[0098] For receiving power: Where: δ represents the overlap between the receiver and the transmitter. ρ s represents the reflectivity of snow particles, c represents the light beam, τ H represents the half-power pulse width, θ j It indicates the part of the laser beam opening angle reflected by snow particles;

[0099] Calculate the sum of the received power of the hard target and all snow particles to get the total received power P R，snow :

[0100] S2-11. Compare the peak value of the reflected power at the point and the size of the hard target at the point. If the peak value of the reflected power is greater than the hard target, the position of the point is updated and then output; otherwise, the position information of the point is directly output;

[0101] S2-12. Use the output spatial position information as simulation data under snowy conditions for training and testing of three-dimensional target detection.

[0102] It is understandable that this point cloud target detection method can more accurately train and test the 3D target detection model by simulating radar point cloud data under snowy conditions. The core is to use the point cloud data collected under sunny conditions as the basis to simulate the existence and influence of snow particles, including the random distribution, occlusion and scattering effects of snow particles, and the reflection power competition between snow particles and hard targets. By calculating the total power received at each point and comparing it with the size of the corresponding hard target, the simulated point cloud data under snowy conditions is finally obtained, which enables the model to better adapt to the real snowy environment and improve the accuracy of target detection.

[0103] In one embodiment, step S3 includes:

[0104] S3-1. Input the rendered point cloud data.

[0105] S3-2. Convert the point cloud data into a spherical projection image to enhance the geometric feature expression of the point cloud.

[0106] S3-3. Use KNN convolution to construct the neighborhood features of the spherical projection image, capture the relationship between each point in the spatial neighborhood, and construct the local geometric structure features of the spherical projection image.

[0107] S3-4. Capture time series information from the time dimension, combine the spatial features of adjacent points, construct local spatial features, combine the spatial features of adjacent points, and use the integration of time-space features to enhance the resolution of severe weather noise.

[0108] S3-5. Convolution operation, combined with residual block and feature reconstruction strategies (such as average pooling, pixel rearrangement and discard rate optimization), can accurately remove point cloud noise under bad weather conditions.

[0109] S3-6. Extract valid points and reconstruct 3D point cloud data, and convert the two-dimensional coordinates of the spherical projection back into three-dimensional point cloud data.

[0110] S3-7. Output the denoised point cloud data for subsequent target detection tasks.

[0111] It is understandable that the rendered point cloud data is first converted into a spherical projection image to enhance the geometric feature expression of the point cloud. Subsequently, the relationship between local points in the point cloud data is captured through the neighborhood feature construction method based on KNN convolution, and the dynamic information of the time dimension is combined with the spatial features of adjacent points to generate a more complete local space-time feature to help the algorithm understand the point cloud data more comprehensively. On this basis, multi-resolution convolution and feature enhancement strategies are used to remove noise from the point cloud data, accurately eliminate environmental interference such as particle noise in severe weather, and improve the quality of point cloud data. Finally, valid points are extracted and 3D point cloud data is reconstructed to make target detection more accurate and reliable.

[0112] In one embodiment, step S4 includes:

[0113] S4-1. Perform multi-scale local feature extraction and modeling on the denoised point cloud data.

[0114] S4-2. Perform generator-discriminator adversarial training on the modeled point cloud data.

[0115] S4-3. Perform high-quality processing on the trained point cloud data by designing multiple loss functions.

[0116] It can be understood that through multi-scale local feature extraction and modeling, the fine-grained details in the point cloud data can be accurately captured, and the generator-discriminator adversarial training technology can be used to enhance the quality and feature expression of the point cloud data. Finally, through the design of multiple loss functions, the trained point cloud data can be refined, thereby improving the accuracy and robustness of point cloud target detection.

[0117] In one embodiment, step S5 includes:

[0118] S5-1. Perform voxel processing on the point cloud data processed by multiple loss functions to divide it into multiple small voxels.

[0119] S5-2. Feature encoding is performed on the points in each voxel by means of a multi-layer perceptron to obtain voxel features.

[0120] S5-3. Perform convolution operations on voxel features through 3D convolutional neural networks to extract high-level semantic features.

[0121] S5-4. Using the 3D feature map, generate 3D candidate boxes through the region proposal network.

[0122] S5-5. Select some key points from the point cloud data processed by multiple loss functions, pool relevant features from the voxel features around the key points, aggregate the voxel features to the key points, and obtain key point level features.

[0123] S5-6. Fuse high-level semantic features with key point-level features, and use the fused features to refine and classify the 3D candidate boxes, determine the types of target objects contained therein, and finally obtain an accurate 3D target detection box.

[0124] According to a second aspect of an embodiment of the present disclosure, a point cloud target detection device based on snow and fog severe weather is provided, comprising:

[0125] An acquisition module, configured to acquire point cloud data collected by a laser radar;

[0126] a rendering module configured to perform severe weather rendering on the point cloud data to generate a simulated severe weather point cloud;

[0127] A denoising module configured to remove and optimize severe weather noise from rendered point cloud data using a deep learning approach;

[0128] The data completion module is configured to input the denoised point cloud data into the upsampling network to complete the loss of valid points caused by bad weather rendering;

[0129] The target detection module is configured to input the point cloud data processed by the upsampling network into the three-dimensional target detection model, perform point cloud target detection, generate target detection prediction results, and complete the final recognition and positioning.

[0130] In summary, the point cloud target detection method and device based on severe weather rendering and upsampling proposed in the present invention uses the data set constructed by the point cloud rendering severe weather algorithm as the target of analysis, and uses the currently advanced target detection algorithm for verification and comparison, with high correctness and accuracy. The data set rendered using the point cloud rendering severe weather algorithm can effectively achieve the point cloud distortion effect under real weather conditions, and the severe weather noise removal algorithm proposed in the present invention can more effectively remove weather noise. At the same time, the proposed point cloud upsampling network can effectively complete the effective points of the object lost due to severe weather. The present invention provides effective technical support for improving three-dimensional detection performance under severe weather conditions.

[0131] 1. A bad weather rendering algorithm was designed during the 3D object detection model training phase.

[0132] 2. The severe weather noise removal algorithm designed during the model testing phase effectively removes severe weather particle noise.

[0133] 3. The proposed point cloud upsampling network can effectively complete the valid points of objects lost due to bad weather.

[0134] 4. The system proposed in the present invention can improve the performance of the three-dimensional detection model under real severe weather conditions.

[0135] The present invention is described by way of embodiments, and those skilled in the art will appreciate that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application belong to the protection scope of the present invention.

Claims

1. A point cloud target detection method based on severe weather rendering and upsampling, characterized in that: include: S1. Obtain point cloud data collected by the laser radar; S2. Perform bad weather rendering on the point cloud data to generate a simulated bad weather point cloud; S3. Remove and optimize the bad weather noise from the rendered point cloud data using deep learning methods; S4. Input the denoised point cloud data into the upsampling network to complete the loss of valid points caused by bad weather rendering; S5. Input the point cloud data processed by the upsampling network into the 3D target detection model, perform point cloud target detection, generate target detection prediction results, and complete the final recognition and positioning.

2. The point cloud target detection method according to claim 1, characterized in that: In step S1, 60% of the snow and fog severe weather data set constructed by rendering is used as a training set and 40% is used as a test set.

3. The point cloud target detection method according to claim 2, characterized in that: When rendering foggy weather based on point cloud, step S2 includes: S2-1. Input the point cloud data collected by the LiDAR under a clear sky, where each point contains the spatial position (x, y, z); S2-2. Calculate the original position R0 of each point: S2-3. Calculate the hard target item of the point using R0 and the attenuation coefficient of the fog; S2-4. Calculate the soft target item using R0, the initial scattering power of the fog, the backscattering system of the fog and the distance of the current signal propagation; S2-5. Compare and if the soft target item of the point is greater than the hard target item, update the spatial position of the point and then output it; otherwise, directly output the position information of the point; S2-6. Use the output spatial position information as simulation data under foggy conditions for training and testing of three-dimensional target detection.

4. The point cloud target detection method according to claim 3, characterized in that: In step S2-3, the hard target item The formula is: Where: α represents the attenuation coefficient of fog, P R，clear (R) is the received signal power in clear weather; In step S2-4, the soft target item The formula is: Where: CAP0 represents the initial scattering power of the fog, β represents the backscattering system of the fog, and R represents the distance of current signal propagation.

5. The point cloud target detection method according to claim 4, characterized in that: In step S2-4, the formula for updating the spatial position of the point is: x′=s·n·x, y′=s·n·y, z′=s·n·z, otherwise the point is directly output, where: s is the scaling factor and n is the noise factor.

6. The point cloud target detection method according to claim 2, characterized in that: When rendering snow weather based on point cloud, step S2 includes: S2-7. Input the point cloud data collected by the radar under a clear sky, where each point contains the spatial position (x, y, z); S2-8. Generate snow particles based on the snowfall rate and the parameters of the radar sensor, and the snow particles are distributed around the radar sensor and conform to a random distribution; S2-9. Calculate the original distance R0 from the point in the scene scanned by the laser radar to the laser radar sensor, determine the number of snow particles intersecting the light beam on the scanning path, and gradually calculate the snow particle occlusion and scattering effects; S2-10. Calculate the hard target at each point, calculate the sum of the received power of the hard target at the point and all snow particles in the current beam path at the point, obtain the total received power, take the maximum value of the received power at the point, and determine the reflected power and the corresponding position according to the maximum value of the received power; Calculating the received power of a hard target Where: CAP0 represents the initial scattering power of fog, α is the incident angle, and ·ρ0 represents the reflectivity of the target object; For receiving power: Where: δ represents the overlap between the receiver and the transmitter. ρ s represents the reflectivity of snow particles, c represents the light beam, τ H represents the half-power pulse width, θ j It indicates the part of the laser beam opening angle reflected by snow particles; Calculate the sum of the received power of the hard target and all snow particles to get the total received power P R，snow : S2-11. Compare the peak value of the reflected power at the point and the size of the hard target at the point. If the peak value of the reflected power is greater than the hard target, the position of the point is updated and then output; otherwise, the position information of the point is directly output; S2-12. Use the output spatial position information as simulation data under snowy conditions for training and testing of three-dimensional target detection.

7. The point cloud target detection method according to any one of claims 1 to 6, characterized in that: Step S3 includes: S3-1. Input the rendered point cloud data; S3-2. Convert the point cloud data into a spherical projection image to enhance the spatial feature expression of the point cloud; S3-3. Use the neighborhood feature construction method based on KNN convolution to extract the local geometric features of the spherical projection image; S3-4. Capture the temporal information of the point cloud from the time dimension, combine the spatial features of adjacent points, and generate local high-dimensional features; S3-5. Through multi-resolution convolution, feature enhancement and noise reduction strategies, severe weather noise is removed with high precision; S3-6. Extract valid points and reconstruct 3D point cloud data, significantly improving the integrity of the point cloud and the accuracy and reliability of target detection. S3-7. Output the denoised point cloud data.

8. The point cloud target detection method according to claim 7, characterized in that: Step S4 includes: S4-1. Perform multi-scale local feature extraction and modeling on the denoised point cloud data; S4-2. Perform generator-discriminator adversarial training on the modeled point cloud data; S4-3. Perform high-quality processing on the trained point cloud data by designing multiple loss functions.

9. The point cloud target detection method according to claim 8, characterized in that: Step S5 includes: S5-1. voxelize the point cloud data processed by the multiple loss functions to divide it into multiple small voxels; S5-2. Feature encoding of points in each voxel is performed by means of a multi-layer perceptron to obtain voxel features; S5-3. Perform convolution operation on voxel features through 3D convolutional neural network to extract high-level semantic features; S5-4. Generate 3D candidate boxes through region proposal network using 3D feature map; S5-5. Select some key points from the point cloud data processed by the multiple loss functions, pool relevant features from the voxel features around the key points, aggregate the voxel features to the key points, and obtain key point level features; S5-6. Fuse high-level semantic features with key point-level features, and use the fused features to refine and classify the 3D candidate boxes, determine the types of target objects contained therein, and finally obtain an accurate 3D target detection box.

10. A point cloud target detection device based on severe weather rendering and upsampling, characterized in that: include: An acquisition module, configured to acquire point cloud data collected by a laser radar; a rendering module configured to perform severe weather rendering on the point cloud data to generate a simulated severe weather point cloud; A denoising module configured to remove and optimize severe weather noise from rendered point cloud data using a deep learning approach; The data completion module is configured to input the denoised point cloud data into the upsampling network to complete the loss of valid points caused by bad weather rendering; The target detection module is configured to input the point cloud data processed by the upsampling network into the three-dimensional target detection model, perform point cloud target detection, generate target detection prediction results, and complete the final recognition and positioning.

Images